An Efficient Data-Dependence Profiler for Sequential and Parallel Programs Zhen Li, Ali Jannesari, and Felix Wolf German Research School for Simulation Sciences, 52062 Aachen, Germany Technische Universitat¨ Darmstadt, 64289 Darmstadt, Germany fz.li, [email protected], [email protected] Abstract—Extracting data dependences from programs data dependences without executing the program. Although serves as the foundation of many program analysis and they are fast and even allow fully automatic parallelization transformation methods, including automatic parallelization, in some cases [13], [14], they lack the ability to track runtime scheduling, and performance tuning. To obtain data dependences, more and more related tools are adopting profil- dynamically allocated memory, pointers, and dynamically ing approaches because they can track dynamically allocated calculated array indices. This usually makes their assessment memory, pointers, and array indices. However, dependence pessimistic, limiting their practical applicability. In contrast, profiling suffers from high runtime and space overhead. To dynamic dependence profiling captures only those depen- lower the overhead, earlier dependence profiling techniques dences that actually occur at runtime. Although dependence exploit features of the specific program analyses they are designed for. As a result, every program analysis tool in need profiling is inherently input sensitive, the results are still of data-dependence information requires its own customized useful in many situations, which is why such profiling profiler. In this paper, we present an efficient and at the forms the basis of many program analysis tools [2], [5], same time generic data-dependence profiler that can be used [6]. Moreover, input sensitivity can be addressed by running as a uniform basis for different dependence-based program the target program with changing inputs and computing the analyses. Its lock-free parallel design reduces the runtime overhead to around 86× on average. Moreover, signature-based union of all collected dependences. memory management adjusts space requirements to practical However, a serious limitation of data-dependence profiling needs. Finally, to support analyses and tuning approaches for is high runtime overhead in terms of both time and space. parallel programs such as communication pattern detection, The former may significantly prolong the analysis, some- our profiler produces detailed dependence records not only for times requiring an entire night [15]. The latter may prevent sequential but also for multi-threaded code. the analysis completely [16]. This is because dependence Keywords-data dependence, profiling, program analysis, par- profiling requires all memory accesses to be instrumented allelization, parallel programming and records of all accessed memory locations to be kept. To lower the overhead, current profiling approaches limit I. INTRODUCTION their scope to the subset of the dependence information Single-core performance is more or less stagnating. needed for the analysis they have been created for, sacrificing Nonetheless, to speed up a program developers can now generality and, hence, discouraging reuse. Moreover, since exploit the potential of multi-core processors and make it current profilers mainly concentrate on the discovery of par- run in parallel. However, fully utilizing this potential is often allelization opportunities, they only support sequential pro- challenging, especially when the sequential version was writ- grams, although they are also needed for parallel programs. ten by someone else. Unfortunately, in many organizations For example, there may still be unexploited parallelism the latter is more the rule than the exception [1]. Many useful hidden inside a parallel program. Furthermore, knowledge of tools have been proposed to assist programmers in paralleliz- data communication patterns, which are nothing but cross- ing sequential applications and tuning their parallel versions thread dependences, can help identify critical performance more easily. Tools for discovering parallelism [2], [3], [4], bottlenecks. Finally, debugging approaches such as data race [5], [6], [7] identify the most promising parallelization detection can also benefit from data dependence information opportunities. Runtime scheduling frameworks [8], [9], [10], to improve their accuracy. [11] add more parallelism to programs by dispatching code To provide a general foundation for all such analyses, sections in a more effective way. Automatic parallelization we present the first generic data-dependence profiler with tools [12], [13], [14] transform sequential into parallel code practical overhead, capable of supporting a broad range automatically. However, they all have in common the fact of dependence-based program analysis and optimization that they rely on data-dependence information to achieve techniques—both for sequential and parallel programs. To their goals because data dependences can present serious achieve efficiency in time, the profiler is parallelized, taking obstacles to parallelization. advantage of lock-free design [17]. To achieve efficiency Data dependences can be obtained in two main ways: in space, the profiler leverages signatures [18], a concept static and dynamic analysis. Static approaches determine borrowed from transactional memory. Both optimizations are application-oblivious, which is why they do not restrict the 1 1:60 BGN loop profiler’s scope in any way. Our profiler has the following 2 1:60 NOM {RAW 1:60|i} {WAR 1:60|i} 3 {INIT } specific features: * 4 1:63 NOM {RAW 1:59|temp1} {RAW 1:67|temp1} • It collects pair-wise data dependences of all the three 5 1:64 NOM {RAW 1:60|i} types (RAW, WAR, WAW) along with runtime control- 6 1:65 NOM {RAW 1:59|temp1} {RAW 1:67|temp1} flow information 7 {WAR 1:67|temp2} {INIT *} 8 1:66 NOM {RAW 1:59|temp1} {RAW 1:65|temp2} • It is efficient with respect to both time and memory 9 {RAW 1:67|temp1} {INIT *} (average slowdown of only 86×, average memory con- 10 1:67 NOM {RAW 1:65|temp2} {WAR 1:66|temp1} sumption of only 1020 MB for benchmarks from NAS 11 1:70 NOM {RAW 1:67|temp1} {INIT *} and Starbench) 12 1:74 NOM {RAW 1:41|block} 13 1:74 END loop 1200 • It supports both sequential and parallel (i.e., multi- threaded) target programs • It provides detailed information, including source code Figure 1. A fragment of profiled data dependences in a sequential program. location, variable name, and thread ID The remainder of the paper is organized as follows. First, we summarize related work in Section II. Then, we describe tasks through a series of static analyses, including alias/edge our profiling approach for sequential target programs in partitioning, equivalence classification, and thinned static Section III. At this point, emphasis is given to the reduction analysis. According to published results, the slowdown of of space overhead. In Section IV, we describe the efficient these approaches stays close to ours when profiling the parallelization. An extension in support of multi-threaded hottest 20 loops (70× on average using SD3 with 8 threads), target programs is presented in Section V. We evaluate but remains much higher when profiling whole programs the accuracy and performance of the full profiler design in (over 500× on average using multi-slicing with 8 threads). Section VI, while we showcase several applications of our Like SD3 and multi-slicing, we parallelize the data- profiler in Section VII. Finally, we conclude the paper and dependence profiling algorithm instead of customizing it. outline future prospects in Section VIII. Unlike these methods, we profile detailed data dependences and control-flow information for not only sequential but also II. RELATED WORK multi-threaded programs. Furthermore, our parallelization After purely static data-dependence analysis turned out to is achieved through lock-free programming, ensuring good be too conservative in many cases, a range of predominantly performance without loss of generality. dynamic approaches emerged. In previous work, their over- head was reduced either by tailoring the profiling technique III. DATA-DEPENDENCE PROFILING to a specific analysis or by parallelizing it. To explain how we reduce the space overhead via signa- Using dependence profiling, Kremlin [2] determines the tures, we start with the profiling approach that supports only length of the critical path in a given code region. Based on sequential programs. Our profiler, which is implemented this knowledge, it calculates a metric called self-parallelism, in C++11 based on LLVM [20], delivers the following which quantifies the parallelism of the region. Instead of information: pair-wise dependences, Kemlin records only the length of the • pair-wise data dependences critical path. Alchemist [4], a tool that estimates the effec- • source code locations of dependences and the names of tiveness of parallelizing program regions by asynchronously the variables involved executing certain language constructs, profiles dependence • runtime control-flow information distance instead of detailed dependences. Although these We profile detailed pair-wise data dependences because approaches profile data dependences with low overhead, the we want to support as many program analyses as possible. underlying profiling technique has difficulty in supporting Control-flow information is necessary for some program support other program analyses. analyses such as parallelism discovery and code partitioning. There are also approaches that reduce